One line item, most of the growth
The headline was a semiconductor market growing by nearly two-thirds in a year, the fastest in two decades. The substance was that almost all of that growth was a single line: memory, rising from around a quarter of industry revenue to nearly half. Everything else grew at a rate that in any other year would be a good cycle and this year is a rounding error.
None of it is bits. The forecast puts DRAM prices up 125 percent for the year and flash up 234 percent, with meaningful relief not expected until late 2027 and new supply only arriving in volume after that. The analysts’ own summary uses the phrase “storage crisis” rather than “memory crisis”, and the reason is the flash number: it is inflating at close to twice the rate of DRAM, and flash is what data platforms are made of.
The forecast also names the mechanism. Bit supply goes where the margin is, which is AI accelerators; the rest of the market — PCs, handsets, industrial and, by extension, enterprise storage arrays — waits two years for the queue to clear. The analysts call it demand destruction, and they say so plainly.
Why this reaches the warehouse first
An enterprise data platform is one of the largest consumers of flash an organisation owns, and its growth is structural: retention rules, regulatory history, the copies that accumulate around every analytics and AI initiative. A modernisation programme — warehouse refresh, lakehouse migration, a new storage tier for retrieval and model state — was almost certainly costed on prices from a year that no longer exists.
The effect is not a small variance. If the storage component of a business case was sized at last year’s cost per terabyte, and that cost has tripled, the programme’s economics have changed category. Options that were marginal are now negative. Options that assumed “buy more capacity” as the easy path now face a two-year queue and a price at the top of the cycle. And the AI use cases that were going to justify the spend each carry a data footprint — indexes, caches, training sets, persisted state — that was probably never sized at all.
What to do with the plan
- Re-run the business case with scenarios. Current prices, a further rise through mid-2027, and the analysts’ moderating curve after that. A plan that only works under one of them is not a plan.
- Rank AI use cases by data footprint as well as value. A use case that needs a full copy of the warehouse re-indexed is a different proposition, in a storage crisis, from one that reads through a governed connector. Footprint is now a first-order selection criterion.
- Find capacity before buying it. Copies, silos, replication overhead and idle headroom are the cheapest storage available this year. The audit costs weeks and usually recovers a meaningful fraction of the planned purchase.
- Take the analysts’ own advice on contracts. Be cautious about supply agreements with unfavourable terms extending beyond 2027. Decoded, that is an analyst house telling its buy-side clients not to lock in at the top. Multi-year subscriptions offered at a discount right now are priced against today’s list, not 2028’s.
- Treat the forecast as a forecast. The figures above are everything that is public. The supply assumptions underneath a 234 percent flash-price call sit inside client-only research, and nobody outside the subscription can inspect them. Plan for the direction; do not build a procurement schedule on the decimal places.
The uncomfortable conclusion
For most organisations the honest reading is that the next eighteen months are the wrong time to expand the data estate on capex, and the right time to understand it — to know what is stored, what is duplicated, what each AI initiative will actually add, and where the governed platform can serve models without new copies. That work has been worth doing for a decade. A storage crisis merely removes the excuse for not doing it.
The analysts describe the cycle as profound but not perennial. Prices will ease when bits arrive in volume. The organisations that come out ahead will be the ones whose data platforms are leaner and better understood when they do — and who did not sign a five-year agreement at the peak to get there.
